²ÝÁñÂÛ̳

Why did you join the IAForTheScience Cofund program?

Alessandro : During the last six months of my Master's degree, I dedicated my time to preparing my thesis, focusing on AI-based algorithms for developmental biology. I worked closely with researchers at the Center for Life NanoScience at the Italian Institute of Technology in Rome. This experience was pivotal, solidifying my decision to pursue research in this field with a PhD. It was also during this period that I discovered the opportunity to apply for the prestigious MSCA COFUND PhD program "AI4theSciences" at PSL University in Paris. The PhD project aligned perfectly with my skills and interests, combining artificial intelligence with theoretical physics and developmental biology. I applied, was shortlisted for the interviews, successfully passed all stages of the selection process, and ultimately was selected for the COFUND program.

Linnea : I first became interested in AI after taking courses in Machine Learning and Image Processing at Imperial, where I learned mainly how to analyse biological images. This motivated me to take a year out of my degree to work in Machine Learning and Data Science, as Technology Consulting Analyst at Accenture in 2019/20. There I learned about AI and DataScience for a wider range of applications from baggage transport to analysing the language in Tweets. I was excited by the possibilities of AI and wanted to research cutting edge applications.  I returned to Imperial and wrote my masters thesis on Biomimetic Layers for Convolutional Neural Networks, showing that adding biologically inspired elements to artificial neural networks can improve their performance.
After my masters I came across the AIforTheSciences programme, in particular the project "Language Acquisition in Brains and Algorithms'' and was thrilled by the idea of using AI to study human intelligence, in particular our capacity for language. The combination of a stimulating project and all the cultural offerings of life in Paris were too much to refuse.

Can you explain your research work at PSL?

Alessandro : At PSL, my research focuses on integrating AI and physics to develop new computational algorithms for studying complex cell systems. Biological tissues are intricate systems composed of numerous interacting cells that exhibit sophisticated collective behaviours. These cells are dynamic - they can move, change neighbours, divide, and die. By applying AI techniques, I extract physical information from images provided by collaborating biologists to formulate meaningful physical equations. These equations help explain the emergent phenomena observed when cells interact to form tissues. Understanding these interactions is crucial for advancing developmental biology and has significant implications for medical research, including regenerative medicine.

Linnea : As mentioned above, my PhD project is entitled "Language Acquisition in Brains and Algorithms". I study how language is represented in the brains of children of different ages, and how language is represented in artificial neural networks as they learn.  Concretely, I work with a hospital in Paris to record how the brains of children of different ages react as they listen to an audiobook, then I give the same audiobook to an AI model and see if its activations can predict the activations in the brain.  So far I have collected data on over 50 participants, making this the first and largest study of natural language using this recording method in children to date. On the AI side I have trained 50 GPT-2 models, during an internship at Meta AI, and investigated how they learn. The comparison between real neurons in the brain and artificial neurons in AI models offers insights both to the neuroscience and the AI community.
I am working to answer questions such as  "in which areas of the brain do children process language?", "is this processing faster or slower than adults?" and " do AI models learn grammatical structures in the same order as children?".

What are the current major challenges/issues of AI in your field of research?

Alessandro : One of the major challenges in applying AI-based computer vision algorithms to study biological cell systems is the difficulty in obtaining large, high-quality datasets for training models. Economic and temporal constraints, as well as technical limitations related to instrumentation and sample fragility, often restrict the availability of extensive data. This makes developing effective AI-based computer vision techniques particularly challenging. The key issue is designing self-supervised models - since labeled data are rarely available - that are sophisticated enough to capture intricate structures in images but not so complex that they require large datasets for training. Balancing these requirements is crucial for advancing research in this field.

Linnea : A major challenge in comparing the representations of language in AI to those in the brain is the difference in the quantity of input data that AI models require to learn compared to humans. While humans can learn from thousands of examples, AI systems require millions of written words to reach a similar ability. In addition most language models take text input, while children learn language from speech, which makes their representations difficult to compare sensibly.  One solution to these problems is using self-supervised audio-input models. One such model, wav2vec 2.0, can achieve impressive missing token prediction after only 53k hours of audio input, which is comparable to the amount of input a child receives.
In order to have models that can learn in the data efficient way that humans do we need to further investigate in which ways our current models are similar and different to the human brain.

The competition was an excellent opportunity to receive feedback on my work from peers and experts, and to network with other young scientists.

On June 10 and 11, you took part in the EELISA scientific competition for students. What does it involve? Why did you take part?

The two PhD students describe the competition as an opportunity for bachelor's, master's and PhD students to present their research projects and innovative solutions to scientific and technological challenges. Each participant was asked to submit a summary of his or her research, after which the selected students provided a detailed 20-page report. The final score was based on a 15-minute presentation and a 5-minute question-and-answer session in front of the jury and audience.

Alessandro : On June 10 and 11, I participated in the EELISA scientific student competition in Budapest, Hungary. This competition involves Bachelor’s, Master’s, and PhD students presenting their research projects and innovative solutions to scientific and technological challenges. Participants submit an abstract of their work, and selected students provide a detailed 20-page report, which contributes to the initial evaluation. The final score is based on a 15-minute presentation and a 5-minute Q&A session in front of a panel of experts. This competition was a very important experience during the final phase of my PhD. I now prepare to wrap up my project and defend my thesis, and this competition provided an excellent opportunity to receive feedback from peers and experts on my work, as well as to network with other young scientists.

Linnea : The EELISA scientific competition took place in person in Budapest, Hungary from June 10-11. I submitted a report on the main study of my thesis for which I have been collecting data and analysing for the past two years. As I will submit a paper on this project to a top tier journal in the next several weeks, the competition was the perfect chance to get feedback from a panel of external judges before journal submission. This was a great opportunity to polish a presentation on the project, and to consider how I could explain the details of my study to a wider audience who may not be familiar with the techniques used in my subfield.

How did the conference go? Do you have a memorable anecdote to tell? What was your experience?

Alessandro : I presented my PhD research in the "Chemical and Bioengineering" section, and was very honored to be awarded first prize in this category. The competition was a memorable and enriching professional and social experience. On the day of the presentation, I was quite stressed beforehand. After the competition, a colleague and I decided to de-stress at the Central Market in Budapest, where we enjoyed a late and hearty lunch of traditional Hungarian food. In the evening, social events were organised for participants to get to know each other and explore the iconic historical areas of the city, Buda and Pest, separated by the Danube River.

Linnea : The conference went well - I was happy to be awarded third place in my category, Medicine. I found the students friendly and outgoing as well as very international. It was enjoyable to learn from scientists at different stages in their careers, from first year of Bachelor up to PhD students.
Several social events were organised to help everyone mix and see the famous Buda castle, the beautiful parliament buildings and the scenic town centre.

What's the next step in your career?

Alessandro : As I mentioned, I am in my third year of my PhD and have just received an extension to my contract to finalise all the projects I have been working on. I plan to submit three different papers to top-tier journals next year, as well as write and defend my thesis by next summer. Looking ahead, I aim to continue working in research on similar topics, ideally within the research department of a company that provides access to the computational resources necessary for cutting-edge AI. I hope to leverage my skills in integrating AI and theoretical physics to explain complex systems in physical terms, whether in biology or other socially relevant domains where individual elements interact to create fascinating emergent phenomena.

Linnea : I am currently preparing for my PhD defence which is scheduled for December, 2024. Afterwards I have secured funding to stay for an additional year in the same lab at PSL as a postdoc. There are a lot of projects I would like to finish and analyses I would like to wrap up on the unique dataset I have collected during my PhD, so I'm happy to have a postdoc year in which to do this. Afterwards I would like to continue working in the intersection of AI and neuroscience, ideally in a research department of a company with the computation resources required for cutting edge AI.

 
Both PhD students say they want to continue working in research on similar topics, ideally in the research department of a company with access to the computing resources needed for cutting-edge AI.